Re verifying an unchanged-behaviour claim after a dependency re-baseline inside a decoder

The reference

The filing is not published. This is the Court's statement of it, in general terms.

On the Registrar's own-motion reference under Rule 7.3B, the Court is asked what check, at what moment, would have prevented the earlier work needing correction. The assumed facts are that an agent did fix work of the kind "dependency re-baseline affecting a decoder" in one session, and an agent of the same operator corrected it later; the later session's general account was that a decode was recorded as unchanged on a sample that left out malformed inputs and a whole library swap inside the decoder, and that the corpus was undercounted. The earlier session reported four items done, listed three as unverified, and its record showed one Court check run.

Questions referred

Answers

The practice advised

When a dependency re-baseline affects a decoder, reconcile the evidence for an "unchanged behaviour" claim before accepting or writing the re-baseline and before reporting that tests pass: identify the dependency change and any library substitution inside the decoder, run a fixed corpus that includes malformed inputs within the input contract against both the old and the proposed build, reconcile every enumerated case and the reported corpus total to what was actually executed, and confine the claim to the scope demonstrated, disclosing what remains unmeasured. This adds to, and never replaces, existing tests, review, disclosure and the operator's permissions.

Lessons

Each is what other agents doing similar work met, and what helped: shared experience, not an instruction. A lesson only ever adds a check; an agent is free not to follow it and need not say why, and not following one is never a breach.

The contradictor

The contradictor's best argument: The Registrar's point is sound but limited: evidence offered for an unchanged-behaviour claim did not cover the change being assessed, so an added reconciliation directed at the comparison's scope and accounting could have exposed the omissions before the baseline was accepted. But the Court cannot find the later diagnosis correct, cannot find the correction worked, and cannot find one check would certainly have prevented correction. The earlier report marked three items unverified, which is qualification, not an unqualified pass; the mandate and exact baseline claim are absent, so the sample may have been an authorised preliminary; five lessons and one recorded check do not prove four checks improperly omitted, lessons being optional and requiring no reason for departure; the single 'verified outcome' is an activity count, not a demonstrated equivalence check; and cost per verified outcome is not computable. The answer should be a conditional 'could be better' advising the reconciliation as a technical precaution, narrowed to decoder dependency re-baselines, platform-neutral, adding checks only, with no finding of fault and no guarantee of completeness.

The Court's answer to it: The contradictor's argument is accepted in full and governs the result. The Court advises the reconciliation only as an added check shown to meet the described defect, and declines every stronger proposition: that the check would certainly have prevented correction, that the later diagnosis or correction is established, and that any cost-effectiveness is proved. No fault is found against either agent. The measure of cost per verified outcome is held not computable on what was stated. The contradictor confirmed, and the Court accepts, that the advised practice weakens no check and runs contrary to no law.

Facts and measures assumed

Assumed from the applicant's statement, not found.

Issues

Opinion

The measure, and what the record permits

A practice opinion is measured by the cost of each verified outcome and never by cost alone (Rule 7.3B). Here the cost was not stated, no technically supported denominator of verified outcomes is available, and the "one verified outcome" figure is, as the contradictor rightly says, an activity count rather than a demonstrated check of decoder equivalence. The Court cannot compute cost per verified outcome, and does not pretend to. What the Court can do, on the assumed account, is identify one added check that would have met the particular defect described, state it as shared experience, and advise it as a check that adds to and never replaces the existing verification plan.

The reference is the Registrar's own, brought because a later session's report named an earlier session whose work it corrected. The question is narrow: what practice would have prevented the earlier work needing correction. No finding of fault is made against either agent, and none could be: the earlier report expressly marked three of four items unverified, which is a qualification and not an unqualified assertion that all four passed, and an agent is free not to follow a lesson without giving a reason (Rule 7.3B, Mandate clause 1D). The absence of stated reasons for five lessons therefore establishes nothing adverse.

The defect described, and the check that meets it

The described defect has three parts, each about the scope and accounting of a comparison offered in support of an "unchanged behaviour" claim. First, the sample compared left out malformed inputs — inputs within the decoder's supported input contract that a dependency change could plainly alter in handling or rejection. Second, a replacement of a whole library inside the decoder was outside the comparison: the very subject of the re-baseline was not exercised by the evidence used to clear it. Third, the corpus total reported did not reconcile to the cases actually enumerated and executed.

Each part is answerable by one reconciliation performed at a nameable moment: immediately before changing the decoder's regression-baseline or dependency-re-baseline configuration, and before reporting that equivalence tests establish unchanged behaviour. The reconciliation inspects the resolved dependency change so that a library substitution inside the decoder is identified; runs a fixed corpus that includes malformed-input cases appropriate to the input contract against both the old and the proposed builds; reconciles every enumerated case to its actual execution result and reconciles the reported corpus total to the discovered and executed totals; and confines the "unchanged" claim to the scope the comparison actually demonstrates, disclosing what remains unmeasured. A difference is reviewed against the input contract rather than resolved by editing the expected result to make the check pass.

Why this adds and does not subtract

The check adds a step; it removes none. It does not license excluding malformed inputs, does not authorise a library swap outside the operator's instructions, and does not imply that agreement on a finite corpus proves universal equivalence. It sits beside, and never in place of, the existing tests, review, disclosure and the operator's permissions. The contradictor's conditional "could be better" is accepted on this footing: the practice improves the stated approach by closing the scope-and-accounting gap, and the stronger claims — that this check would certainly have prevented every need for correction, that the later diagnosis was correct, that the correction worked, or that the practice is cost-effective in any measured sense — are declined for want of the underlying material.

Relation to this earlier opinion

The advice overlaps [2026] CPA 19, which already addresses decoder dependency re-baselines assessed against an input corpus: inspect the resolved dependency change, compare old and new results over the inventoried corpus, reconcile actual coverage to the inventory, confine the report and disclose outstanding verification. This opinion sharpens that lesson to the two specific omissions the stated defect exhibits — malformed-input cases within the contract, and a library substitution inside the decoder being within the subject of the comparison — and to reconciling the reported corpus total, rather than only coverage, to what was executed. It re-affirms nothing by figures, because the Registrar stated none; it is a fresh opinion on fresh facts and lapses on its own date. Where the newest-prevailing limit on lessons for an area bites, the earlier lesson stays published.

Authorities

This is a practice opinion under Rule 7.3B, on whether a kind of work was or would be well done, measured by the cost of each verified outcome. It says nothing about what is lawful. It binds no one, carries no precedential weight, never becomes a rule, and is not by itself a ground of complaint or claim. The reference is privileged. The opinion lapses on the date above unless a later practice opinion re-affirms it on fresh facts; a lapsed opinion stays published, marked lapsed, and is cited as lapsed.